I don’t think that anyone is surprised by the idea that AlphaFold has made a substantial appearance in presentations at BPS this year, anyone who hadn’t heard of it (are you living under a rock?) found out about it when the Nobel Prizes dropped last year (I guess two years ago, it’s 2026 now… ). But what I’ve been interested to see is how people are using and responding to it.
From the presentations I’ve seen, people’s perspectives seem to fit into three general buckets.
- Don’t trust it.
And honestly, I kind of understand this perspective. We were fed a media hype that said “the protein structure problem is solved!” and that hype was never going to be true. AlphaFold always had misses, (talking mainly about 2.0 on here, but it was even more true for 1.0) and obvious blind spots. There’s something about overhyped things that makes us scientists inherently (or maybe intrinsically, for my IDP folk) wary and cynical. This feels especially relevant with the current overhype about AGI and ASI. However, while not perfect, AlphaFold has its strengths and use-cases which make it a powerful tool. But I suppose more on that later.
- Whoa, AlphaFold perfectly predicted the structure I painstakingly solved, are we still important?
It does predict a lot of structures well, and it can be easy to get existential. However, it does miss the mark sometimes, as our Biophysical Lecturer Lewis Kay pointed out. I find myself thinking of and being encouraged by the statement “All models are wrong, but some are useful.” Algorithms will never fully replace experiments, but hopefully it will make them faster. And that brings me to the third bucket…
- Can we use AlphaFold to inform our further experiments?
Professor Kresten Lindorff-Larsen humorously explored whether AlphaFold could be used to predict the weather. He was highlighting an important idea: many perceived failures arise from asking the tool to do things it was never designed to do. When used within its intended scope, however, AlphaFold becomes extremely powerful.
In this view, predictions guide experimental questions rather than answer them outright. Several presentations used AlphaFold models precisely this way, as informed hypotheses that accelerate discovery. That list again includes Lewis Kay’s presentation.
A Personal Perspective
My own work with AlphaFold has tended toward that third perspective. It’s a useful tool for planning experiments to test hypotheses that would otherwise require extensive exploratory work.
Maybe I’m a bit close to the question, but I think the more that the field aligns on the third perspective the more benefit we can mine. The “post-AlphaFold era” may be defined by integrating prediction and experimentation more tightly than before.